The Book That Hides Its Own Arguments

AA dictionary reaches the hand looking like a report on a language rather than a set of decisions about one. Its columns are even, its abbreviations uniform, and nothing in the typography preserves the arguments that ended in this word being entered and that one left out. The Bellwood Dictionary of English, proposed to a London learned society in 1861 and completed seventy years later, ran to 414,000 entries; a reader who trusts the round number may conclude that English contained 414,000 words in 1931. What the figure actually records is narrower: that one editorial board, working from one pile of evidence, judged 414,000 items worth the space a publisher had agreed to pay for. Lexicographers have never pretended otherwise. The difficulty is that the finished object conceals the reasoning, and every quarrel about how dictionaries should be made since has turned on a single question — what kind of evidence a maker is entitled to trust.

BThe method Bellwood's first editor devised has an austere logic. Between 1862 and 1928 some 1,100 volunteers read books assigned to them and copied onto paper slips any sentence in which a word seemed to be doing something worth recording; 5.1 million slips reached the editorial rooms in Clerkenwell. The slips were filed by headword, and an entry was written from the resulting pile: the earliest dated example fixed the word's first known appearance, and later ones traced how its senses had divided. Nothing entered the dictionary that a volunteer had not first found in print and a sub-editor had not then dated. The strength of the procedure is that every claim in a finished entry can be traced to a page and a year. Its weakness follows from the same design, for a slip records what struck a reader as remarkable, and the ordinary is by definition unremarkable. In sixty-six years of reading, the word ‘the’ generated fewer slips than the word ‘quiddity’.

CHelena Vasková, who directs the corpus programme at the Rendle Institute, argues that a dictionary built from slips reports the reading habits of its volunteers as faithfully as it reports the language. Her group assembled a corpus of 1.4 billion words of present-day English — newspapers, transcribed conversation, technical manuals, message boards, fiction — sampled so that each category occupies the share of the whole that surveys of national reading and speaking time assign it. Every occurrence counts once, so frequency rather than salience decides what an editor sees first, and the definite article alone accounts for one word in every sixteen. Run against Bellwood's 1931 text, the corpus showed that 26 per cent of that edition's headwords did not occur even once, while 4,300 items appearing more than fifty times per million had no entry at all. Vasková is careful about what the exercise proves: a corpus is a record of texts that happened to be collected rather than of a language, and she calls her own sampling frame “the best available guess at proportion, and no more than a guess”.

DRobert Innes, general editor of the Bellwood revision until 2018, replies that frequency answers a question lexicography was not asking. A dictionary is opened, he observes, precisely when a reader meets a word that is not familiar, so the entries that matter most are the ones a frequency list ranks lowest. From the online edition's search records he took every query logged in a single month of 2016 and matched it against Vasková's corpus: 71 per cent of the words looked up fell below one occurrence per million. Innes grants two weaknesses. His volunteers were self-selected and read what they already owned, so the slip file leaned heavily towards printed literature, and its coverage of speech was negligible until 1975. He accepts too that search records show what readers fail to understand rather than what they use, and that a month in 2016 may not resemble a month in any other year.

EThe quarrel does not stay on the page. In 2014 an appeal court settled a contract dispute by quoting Bellwood's definition of ‘promptly’, and examination boards in three countries name a dictionary edition that candidates may bring into the hall. Publishers, meanwhile, advertise each revision by the number of words it adds, while the words a revision quietly drops are counted nowhere. Vasková points out, awkwardly for her own position, that no frequency table can tell a court which of two senses a contract intended, because counting establishes what is common and not what is meant. Innes answers that the awkwardness is shared: once a definition is read as though it were law, an editor's silence about how the entry was assembled becomes a claim in itself. Both dislike the way a revision is publicised, and both say that the lexicography which reaches the public is the part that can be summarised in a figure.

FWhat has emerged is a division of labour rather than a verdict. Where a word is common, frequency is the better guide, since a corpus shows in an afternoon which of its senses now dominates and slips cannot. Where a word is rare, the citation file remains the only record of it, and no amount of sampling will supply an example that the sampling never caught. Whether one book can serve the reader who wants the ordinary described and the reader who wants the obscure explained remains open, and no method now in use settles it. What both researchers reject is the assumption that has shaped the argument since 1861: that evidence about a language either describes how it is used or prescribes how it should be, and cannot quietly be doing both.